Hugging Face's logo

Datasets:
SupraBench
/
EU-PMC

Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Size:
100K - 1M
ArXiv:
Tags:
chemistry
supramolecular
host-guest
molecular-recognition
macrocyclic
continued-pretraining
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
EU-PMC / README.md
mtybilly's picture
mtybilly
Add task_categories metadata (supersedes PR)
4bc0889 verified
preview code
|
Raw
History Blame Contribute Delete
6.33 kB
metadata 
license: other
task_categories:
  - text-generation
tags:
  - chemistry
  - supramolecular
  - host-guest
  - molecular-recognition
  - macrocyclic
  - continued-pretraining
size_categories:
  - 100K<n<1M
configs:
  - config_name: default
    data_files:
      - split: raw
        path: data/raw-*.parquet
      - split: filtered
        path: data/filtered-*.parquet

SupraBench

SupraBench is the first benchmark for evaluating large language models on supramolecular host-guest chemistry reasoning. It comprises four fundamental tasks plus an auxiliary vision task, and ships a domain text corpus for domain-adaptive pretraining (DAPT).

Supramolecular chemistry studies non-covalent host-guest assemblies that underpin drug delivery, chemical sensing, and in-vivo toxin sequestration. Designing host-guest systems is slow (days of dry-lab verification per pair); SupraBench probes whether LLMs can reason about these systems directly.

The dataset family

Dataset Task Description
SupraBench/bap Binding Affinity Prediction regress log K_a for a host-guest pair
SupraBench/tbs Top-Binder Selection pick the strongest binder among 4 candidate guests
SupraBench/sid Solvent Identification 6-way solvent classification from structure
SupraBench/hgd Host-Guest Description open-ended QA on host/guest property profiles
SupraBench/EU-PMC Text corpus 16M-token supramolecular corpus for DAPT
SupraBench/Binding-Affinity Comprehensive anchor per-record binding data + host/guest SMILES, 2D, 3D, environment

Each task dataset has a test split (merged records) and a cb7 split (the CB[7] supplement, for add-on evaluation). Each base/fewshot/cot rendering is tagged by the prompt_strategy field.

Dataset statistics

Task # Samples
BAP 2,609
TBS 2,264
SID 2,172
HGD 135

Top-4 hosts (BAP / TBS / SID counts): CB[8] 261/200/571, CB[7] 217/200/217, beta-CD 201/200/264, p-SC4 144/144/225.

SupraCorpus (EU-PMC): 420,950 raw filtered articles -> 133,867 high-precision articles -> ~16M tokens.

Performance report

Main results from the SupraBench paper across the four fundamental tasks (8 LLMs x 3 prompting strategies). Bold = best, italic = second-best per column. Arrows give the optimization direction.

Base

Model BAP MAE down BAP RMSE down TBS ACC up TBS Regret down SID F1 up SID B.Acc up HGD Recall up HGD Prec up HGD F1 up
Qwen3.5-9B 2.491 3.360 0.379 0.930 0.159 0.166 0.040 0.023 0.043
Qwen3.5-27B 1.803 2.503 0.404 0.851 0.225 0.364 0.495 0.072 0.122
Llama3.1-8B 2.699 3.630 0.228 1.281 0.151 0.225 0.266 0.059 0.092
Llama3.1-70B 1.632 2.149 0.338 1.054 0.118 0.254 0.487 0.091 0.152
GPT-5.4-Mini 1.549 2.182 0.428 0.810 0.219 0.274 0.437 0.086 0.137
GPT-5.4-Nano 1.642 2.169 0.411 0.816 0.182 0.347 0.472 0.062 0.107
Gemini-3-Flash 1.248 1.679 0.498 0.647 0.350 0.470 0.506 0.067 0.118
DeepSeek-v4 1.433 1.994 0.461 0.730 0.309 0.381 0.500 0.090 0.141

Few-Shot

Model BAP MAE down BAP RMSE down TBS ACC up TBS Regret down SID F1 up SID B.Acc up HGD Recall up HGD Prec up HGD F1 up
Qwen3.5-9B 3.650 4.820 0.370 0.951 0.154 0.150 0.000 0.022 0.042
Qwen3.5-27B 2.258 3.256 0.392 0.889 0.178 0.257 0.636 0.585 0.580
Llama3.1-8B 5.504 6.940 0.283 1.227 0.142 0.182 0.655 0.369 0.456
Llama3.1-70B 1.774 2.359 0.354 1.026 0.144 0.185 0.631 0.474 0.531
GPT-5.4-Mini 1.958 2.808 0.430 0.824 0.141 0.291 0.542 0.228 0.307
GPT-5.4-Nano 2.176 2.894 0.419 0.819 0.190 0.270 0.532 0.095 0.152
Gemini-3-Flash 1.257 1.702 0.513 0.619 0.389 0.421 0.660 0.364 0.448
DeepSeek-v4 1.618 2.276 0.470 0.713 0.203 0.225 0.720 0.303 0.352

CoT

Model BAP MAE down BAP RMSE down TBS ACC up TBS Regret down SID F1 up SID B.Acc up HGD Recall up HGD Prec up HGD F1 up
Qwen3.5-9B 3.664 4.885 0.382 0.944 0.167 0.197 0.300 0.039 0.068
Qwen3.5-27B 2.438 3.468 0.398 0.898 0.254 0.415 0.526 0.051 0.092
Llama3.1-8B 4.911 6.279 0.293 1.220 0.154 0.153 0.380 0.102 0.144
Llama3.1-70B 1.833 2.512 0.373 0.985 0.106 0.380 0.421 0.055 0.097
GPT-5.4-Mini 2.036 2.887 0.429 0.828 0.220 0.282 0.444 0.080 0.129
GPT-5.4-Nano 2.160 2.881 0.410 0.822 0.174 0.257 0.492 0.056 0.098
Gemini-3-Flash 1.261 1.723 0.510 0.609 0.331 0.432 0.512 0.062 0.110
DeepSeek-v4 1.541 2.183 0.445 0.743 0.307 0.414 0.522 0.080 0.134

Takeaways: frontier proprietary LLMs (Gemini-3-Flash, DeepSeek-v4) lead the quantitative tasks, yet every task leaves substantial headroom; no single prompting strategy is universally best; and CoT amplifies rather than fixes the underlying reasoning gap on binding-affinity prediction.

Sources & license

Binding records are derived from SupraBank (CC-BY-4.0); the text corpus is built from open-access Europe PMC articles subject to each article's individual license; molecular structures use PubChem and OPSIN.

If you use SupraBench, please cite the paper and the upstream data sources.